The default is one agent with many tools
Most systems begin as a single agent holding every tool. It works at first and degrades as the tool count grows: selection accuracy falls, the prompt accumulates conditional instructions, and every change risks every behaviour.
The degradation is gradual, which is why it is usually addressed late — there is no single moment where it breaks, just a slow increase in the rate of odd decisions.
Supervisor
A supervisor delegates to narrow specialists, each with a small tool surface and a clearly described remit. Selection happens twice — which specialist, then which tool — but each decision is made against far fewer options, which is generally more reliable than one decision against many.
The property that matters most is testability. Each specialist can be evaluated independently, with its own fixtures. When a run goes wrong, the hand-off points localise the fault, which is the closest thing to a stack trace that agent systems offer.
The cost is an extra hop, and therefore latency and tokens. For most business systems that trade is worth it.
Swarm
Peer agents handing off to each other without central coordination can be more flexible and is genuinely appealing for open-ended exploration. It is also much harder to reason about: control flow is emergent, termination is a design problem rather than a given, and cost is bounded only by whatever limit you remembered to impose.
Worth choosing deliberately for problems whose shape is genuinely unknown in advance. Worth avoiding for workflows that have a known structure, which is most commercial work.
The topology nobody wants to recommend
A large share of workflows sold as agents are sequences with conditionals. Read a document, extract fields, validate them, take one of three actions. That is a pipeline with an LLM in some steps, and building it as a pipeline makes it cheaper, faster, more reliable and dramatically easier to test.
The honest recommendation, more often than the market suggests, is not to build an agent at all. Reserve agency for problems where the sequence genuinely cannot be known ahead of time — and be suspicious of how often that is claimed.
Whatever you choose, make hand-offs explicit
The most useful structural property across topologies is that transitions are inspectable: what was passed, why, and what came back. Implicit hand-offs through shared mutable state are the agent equivalent of global variables, and they fail the same way — invisibly, at a distance, under concurrency.